Serialize maps without hashes.

runtime/lib/compact_hash.dart

 // Copyright (c) 2015, the Dart project authors.  Please see the AUTHORS file
 // for details. All rights reserved. Use of this source code is governed by a
 // BSD-style license that can be found in the LICENSE file.
 
 import 'dart:typed_data';
 import 'dart:_internal' as internal;
 
 // Hash table with open addressing that separates the index from keys/values.
 
 abstract class _HashFieldBase {
   // Each occupied entry in _index is a fixed-size integer that encodes a pair:
   //   [ hash pattern for key | index of entry in _data ]
   // The hash pattern is based on hashCode, but is guaranteed to be non-zero.
   // The length of _index is always a power of two, and there is always at
   // least one unoccupied entry.
+  // NOTE: When maps are deserialized, their _index and _hashMask is regenerated
+  // lazily by _regenerateIndex.
+  // TODO(koda): Consider also using null _index for tiny, linear-search tables.
   Uint32List _index = new Uint32List(_HashBase._INITIAL_INDEX_SIZE);
 
-  // Cached in-place mask for the hash pattern component. On 32-bit, the top
-  // bits are wasted to avoid Mint allocation.
-  // TODO(koda): Reclaim the bits by making the compiler treat hash patterns
-  // as unsigned words.
-  int _hashMask = internal.is64Bit ?
-      (1 << (32 - _HashBase._INITIAL_INDEX_BITS)) - 1 :
-      (1 << (30 - _HashBase._INITIAL_INDEX_BITS)) - 1;
+  // Cached in-place mask for the hash pattern component. 
+  int _hashMask = _HashBase._indexSizeToHashMask(_HashBase._INITIAL_INDEX_SIZE);
 
   // Fixed-length list of keys (set) or key/value at even/odd indices (map).
   List _data = new List(_HashBase._INITIAL_INDEX_SIZE);
 
   // Length of _data that is used (i.e., keys + values for a map).
   int _usedData = 0;
 
   // Number of deleted keys.
   int _deletedKeys = 0;
 }
@@ skipping 25 matching lines @@
   // The length of _index is twice the number of entries in _data, and both
   // are doubled when _data is full. Thus, _index will have a max load factor
   // of 1/2, which enables one more bit to be used for the hash.
   // TODO(koda): Consider growing _data by factor sqrt(2), twice as often. 
   static const int _INITIAL_INDEX_BITS = 3;
   static const int _INITIAL_INDEX_SIZE = 1 << (_INITIAL_INDEX_BITS + 1);
   
   // Unused and deleted entries are marked by 0 and 1, respectively.
   static const int _UNUSED_PAIR = 0;
   static const int _DELETED_PAIR = 1;
-  
+
+  // On 32-bit, the top bits are wasted to avoid Mint allocation.
+  // TODO(koda): Reclaim the bits by making the compiler treat hash patterns
+  // as unsigned words.
+  static int _indexSizeToHashMask(int indexSize) {
+    int indexBits = indexSize.bitLength - 1;
+    return internal.is64Bit ? (1 << (32 - indexBits)) - 1 :
+                              (1 << (30 - indexBits)) - 1;
+  }
+
   static int _hashPattern(int fullHash, int hashMask, int size) {
     final int maskedHash = fullHash & hashMask;
     // TODO(koda): Consider keeping bit length and use left shift.
     return (maskedHash == 0) ? (size >> 1) : maskedHash * (size >> 1);
   }
 
   // Linear probing.
   static int _firstProbe(int fullHash, int sizeMask) {
     final int i = fullHash & sizeMask;
     // Light, fast shuffle to mitigate bad hashCode (e.g., sequential).
@@ skipping 44 matching lines @@
       // TODO(koda): Consider in-place compaction and more costly CME check.
       _init(_index.length, _hashMask, _data, _usedData);
     } else {
       // TODO(koda): Support 32->64 bit transition (and adjust _hashMask).
       _init(_index.length << 1, _hashMask >> 1, _data, _usedData);
     }
   }
   
   void clear() {
     if (!isEmpty) {
-      _init(_index.length, _hashMask);
+      // Use _data.length, since _index might be null.
+      _init(_data.length, _hashMask);
     }
   }
 
   // Allocate new _index and _data, and optionally copy existing contents.
   void _init(int size, int hashMask, [List oldData, int oldUsed]) {
     assert(size & (size - 1) == 0);
     assert(_HashBase._UNUSED_PAIR == 0);
     _index = new Uint32List(size);
     _hashMask = hashMask;
     _data = new List(size);
     _usedData = 0;
     _deletedKeys = 0;
     if (oldData != null) {
       for (int i = 0; i < oldUsed; i += 2) {
         var key = oldData[i];
         if (!_HashBase._isDeleted(oldData, key)) {
           // TODO(koda): While there are enough hash bits, avoid hashCode calls.
           this[key] = oldData[i + 1];
         }
       }
     }
   }
+
+  int _regenerateIndex() {
+    assert(_index == null);
+    _index = new Uint32List(_data.length);
+    assert(_hashMask == 0);
+    _hashMask = _HashBase._indexSizeToHashMask(_index.length);
+    final int tmpUsed = _usedData;
+    _usedData = 0;
+    for (int i = 0; i < tmpUsed; i += 2) {
+      // TODO(koda): Avoid redundant equality tests and stores into _data.
+      this[_data[i]] = _data[i + 1];
+    }
+    return _index.length;

[siva, 2015/06/01 22:19:36]

instead of checking every caller to this function why not just do
if (_index == null) {
  ......
  ......
}
return _index.length;

[koda, 2015/06/01 23:01:01]

Since we don't do partial inlining (right?), that will either add a function
call to all those operations or blow up their code size.

Do you think this is a premature optimization?

[siva, 2015/06/02 00:13:35]

I don't feel strongly about moving the null check into the function but it does
make the code seem verbose at the call sites.

+  }
   
   void _insert(K key, V value, int hashPattern, int i) {
     if (_usedData == _data.length) {
       _rehash();
       this[key] = value;
     } else {
       assert(1 <= hashPattern && hashPattern <  (1 << 32));
       final int index = _usedData >> 1;
       assert((index & hashPattern) == 0);
       _index[i] = hashPattern | index;
@@ skipping 24 matching lines @@
           }
         }
       }
       i = _HashBase._nextProbe(i, sizeMask);
       pair = _index[i];
     }
     return firstDeleted >= 0 ? -firstDeleted : -i;
   }
   
   void operator[]=(K key, V value) {
-    final int size = _index.length;
+    final int size = (_index == null) ? _regenerateIndex() : _index.length;
     final int sizeMask = size - 1;
     final int fullHash = _hashCode(key);
     final int hashPattern = _HashBase._hashPattern(fullHash, _hashMask, size);
     final int d = _findValueOrInsertPoint(key, fullHash, hashPattern, size);
     if (d > 0) {
       _data[d] = value;
     } else {
       final int i = -d;
       _insert(key, value, hashPattern, i);
     }
   }
   
   V putIfAbsent(K key, V ifAbsent()) {
-    final int size = _index.length;
+    final int size = (_index == null) ? _regenerateIndex() : _index.length;
     final int sizeMask = size - 1;
     final int maxEntries = size >> 1;
     final int fullHash = _hashCode(key);
     final int hashPattern = _HashBase._hashPattern(fullHash, _hashMask, size);
     final int d = _findValueOrInsertPoint(key, fullHash, hashPattern, size);
     if (d > 0) {
       return _data[d];
     }
     // 'ifAbsent' is allowed to modify the map.
     List oldData = _data;
     int oldCheckSum = _checkSum;
     V value = ifAbsent();
     if (_isModifiedSince(oldData, oldCheckSum)) {
       this[key] = value;
     } else {
       final int i = -d;
       _insert(key, value, hashPattern, i);
     }
     return value;
   }
   
   V remove(Object key) {
-    final int size = _index.length;
+    final int size = (_index == null) ? _regenerateIndex() : _index.length;
     final int sizeMask = size - 1;
     final int maxEntries = size >> 1;
     final int fullHash = _hashCode(key);
     final int hashPattern = _HashBase._hashPattern(fullHash, _hashMask, size);
     int i = _HashBase._firstProbe(fullHash, sizeMask);
     int pair = _index[i];
     while (pair != _HashBase._UNUSED_PAIR) {
       if (pair != _HashBase._DELETED_PAIR) {
         final int entry = hashPattern ^ pair;
         if (entry < maxEntries) {
           final int d = entry << 1;
           if (_equals(key, _data[d])) {
             _index[i] = _HashBase._DELETED_PAIR;
             _HashBase._setDeletedAt(_data, d);
             V value = _data[d + 1];
             _HashBase._setDeletedAt(_data, d + 1);
             ++_deletedKeys;
             return value;
           }
         }
       }
       i = _HashBase._nextProbe(i, sizeMask);
       pair = _index[i];
     }
     return null;
   }
   
   // If key is absent, return _data (which is never a value).
   Object _getValueOrData(Object key) {
-    final int size = _index.length;
+    final int size = (_index == null) ? _regenerateIndex() : _index.length;
     final int sizeMask = size - 1;
     final int maxEntries = size >> 1;
     final int fullHash = _hashCode(key);
     final int hashPattern = _HashBase._hashPattern(fullHash, _hashMask, size);
     int i = _HashBase._firstProbe(fullHash, sizeMask);
     int pair = _index[i];
     while (pair != _HashBase._UNUSED_PAIR) {
       if (pair != _HashBase._DELETED_PAIR) {
         final int entry = hashPattern ^ pair;
         if (entry < maxEntries) {
@@ skipping 274 matching lines @@
   E lookup(Object o) => _validKey(o) ? super.lookup(o) : null;
   bool remove(Object o) => _validKey(o) ? super.remove(o) : false;
 
   _CompactLinkedCustomHashSet(this._equality, this._hasher, validKey)
       : _validKey = (validKey != null) ? validKey : new _TypeTest<E>().test;
 
   Set<E> toSet() =>
       new _CompactLinkedCustomHashSet<E>(_equality, _hasher, _validKey)
           ..addAll(this);
 }